GeoWeave-8B achieves the highest open-source final-answer accuracy and process average on GeoVAD-Bench, gaining 25.3 and 30.4 percentage points over its base model¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? GeoWeave-8B reaches 62.6% answer accuracy and 82.9% process average, improving over SenseNova-U1-8B by 25.3 and 30.4 percentage points, with coordinated gains across all four process metrics. → Zhitong Dong 2026 q2 i? GeoWeave-8B outperforms MathCanvas-7B on all four conventional mathematical reasoning benchmarks and achieves higher Math-VR PS and AC than both compared methods. → Zhitong Dong 2026 (2)

Evidence¶

Zhitong Dong 2026¶

Zhitong Dong, Jicai Pan, Yingguo Gao, Jingting Ding, Hao Chen, Jinjie Gu. (2026). Beyond Generation and Accuracy: Diagnosing and Enhancing Visual Chain-of-Thought for Geometry Problem Solving. arXiv preprint. https://arxiv.org/abs/2609.12606 doi:10.48550/arxiv.2609.12606

q2 · i? · design · r2

Main comparison on GeoVAD-Bench under Auto-Aux (Table 7, Section 6.2). The article reports GeoWeave-8B "achieves the highest final-answer accuracy, reaching 62.6%" and 82.9% process average, with per-dimension gains of 5.8, 51.0, 22.0, and 43.0 percentage points.

"Among the open-source models, GeoWeave-8B achieves the highest final-answer accuracy, reaching 62.6%. It improves over the SenseNova-U1-8B baseline by 25.3 percentage points and surpasses every other open-source model listed in Table 7."

Zhitong Dong 2026 (2)¶

Zhitong Dong, Jicai Pan, Yingguo Gao, Jingting Ding, Hao Chen, Jinjie Gu. (2026). Beyond Generation and Accuracy: Diagnosing and Enhancing Visual Chain-of-Thought for Geometry Problem Solving. arXiv preprint. https://arxiv.org/abs/2609.12606

q2 · i? · design · r2

Generalization evaluation on public benchmarks (Table 8, Section 6.3). The article reports GeoWeave-8B "outperforms MathCanvas-7B on all four conventional mathematical reasoning benchmarks" and higher Math-VR PS (64.7) and AC (41.4) than both compared methods.

"GeoWeave-8B outperforms MathCanvas-7B on all four conventional mathematical reasoning benchmarks. On Math-VR, GeoWeave-8B also achieves higher PS and AC than both MathCanvas-7B and CodePlot-CoT-32B."

Discussion¶

Related Claims¶

• Reasoning with ground-truth auxiliary diagrams consistently outperforms reasoning from the original diagram alone on GeoVAD-Bench — related
• Step-level credit assignment added to Interleave-RL raises the reasoning score from 46.3% to 66.3% on GeoVAD-Bench — related
• Correct final answers are associated with stronger performance on all four process-level metrics, with reasoning-process correctness showing the clearest separation — related